The Open Safe AI Alliance expands at Black Hat with SAFE pointers and open frameworks, providing IT leaders inspectable instruments, decrease token prices, and a transparent roadmap for safe enterprise AI.
Applied sciences scale a lot sooner when they’re open and interoperable. This has been true for networking, VoIP, cellular, and nearly each different know-how shift over the previous few a long time. The Open Safe AI Alliance is a big step towards making AI protection open, inspectable, and enterprise-ready, and it arrives at a second when IT groups try to show pilot tasks into manufacturing worth.
Why an open AI protection stack issues
Open supply has lengthy been the spine of recent infrastructure, spanning cloud and compute platforms, and cybersecurity is amongst its greatest beneficiaries. The Open Safe AI Alliance (OSAA) applies that very same mannequin to AI safety by committing to shared open fashions, agent harnesses, instruments, and information that defenders can examine, prolong, and run on their very own infrastructure.
The core argument behind OSAA is that if the subsequent technology of cyberattacks is AI-powered, the subsequent technology of cyber protection should be too, which means defenders want techniques they will really see into. In a current incident on Hugging Face, closed AI instruments had been blocked, stopping important forensic evaluation and forcing the workforce to depend on an open-weight GLM 5.2 mannequin operating by itself infrastructure to research greater than 17,000 actions and include the intrusion.
That episode sheds mild on a brand new fact for CISOs and safety operations leaders: for those who can’t examine and adapt your AI defenses within the second of a breach, you’re blind.
Who’s within the alliance and why it’s completely different
OSAA brings collectively a broad roster spanning cloud, cybersecurity, enterprise software program, and open-source foundations, amongst others. For enterprises, this issues as a result of it indicators that the AI protection stack gained’t be a single-vendor vertical product however an ecosystem anchored in open requirements and parts.
A number of of the founding contributions spotlight how deep this stack can go:
- HPE is contributing to SPIFFE/SPIRE, advancing zero-trust identification requirements that may cryptographically confirm AI brokers and companies earlier than they work together with enterprise sources.
- Hugging Face has launched Safetensors, a safer model-weight format that forestalls distant code execution and will increase transparency in how fashions are saved and loaded.
- IBM and Crimson Hat’s Lightwell extends safety throughout the open-source provide chain with digitally signed patches, serving to shut gaps within the software program layer that AI techniques depend on.
- Microsoft’s MDASH is a multi-model, agentic scanning harness that makes use of specialised brokers to find, debate, and show the existence of exploitable bugs, relatively than counting on a single opaque scanner.
These parts type an rising “open protection stack” for AI brokers that spans identification, isolation, secure mannequin codecs, multi-model scanning, and safe improvement workflows. For IT leaders, this issues as a result of it guarantees constant controls throughout a multi-vendor atmosphere, relatively than brittle level instruments that don’t interoperate.
Nvidia, Cisco and the function of agent harnesses
Nvidia is backing the alliance not solely with fashions and weights but additionally with open agent-harness analysis by way of its Nvidia Labs Object-Oriented Agent (NOOA) framework. NOOA begins with a key perception for practitioners: mannequin selection is barely half the equation; the harness, the structure across the mannequin, can account for double-digit swings in benchmark efficiency and enormous variations in token value, even when the underlying mannequin stays unchanged.
In NOOA, an agent is represented as a single Python class. Its strategies outline capabilities, fields outline state, docstrings function prompts, and kind annotations implement contracts. That design makes agent improvement feel and appear like conventional software program engineering: brokers may be diffed, code-reviewed, unit-tested, and traced with the identical instruments groups already use for different crucial techniques.
For enterprises, it is a delicate however essential shift because it turns agentic AI from a “black field workflow graph” right into a testable, auditable code floor that safety and platform groups can govern.
NOOA identifies six harness capabilities that materially have an effect on mannequin efficiency:
- Typed enter and output, so calls have validated arguments and returns as a substitute of free textual content.
- Go-by-reference semantics, the place the mannequin operates on dwell Python objects and bounded previews as a substitute of serialized dumps pushed into the context window.
- Code-as-action, which means the mannequin writes Python that invokes strategies and management circulation straight.
- Programmable loop engineering, with orchestration loops expressed as unusual Python that people and fashions can modify.
- Express object state, holding sturdy typed state on the agent object as a substitute of solely in dialog historical past.
- Mannequin-callable harness APIs for context blocks and occasion historical past, letting the agent examine and handle its personal working set.
For IT execs, harness engineering is essential as a result of it improves effectivity. In SWE-bench, NOOA achieves parity or higher accuracy with roughly half the tokens of comparability harnesses, thanks largely to pass-by-reference and the flexibility to keep away from pricey context-compaction pipelines. For enterprises wrestling with LLM value curves, that mixture of upper accuracy and decrease consumption straight accelerates time to worth.
Cisco’s presence within the alliance builds on its broader partnership with Nvidia throughout networking, information heart AI materials, and safe AI factories. Cisco has already dedicated to integrating applied sciences similar to Silicon One and specialised SuperNICs into AI networking platforms and is advancing safe AI infrastructure from core to edge by way of choices similar to Cisco Safe AI Manufacturing facility with Nvidia.
As these networking and safety layers intersect with open harnesses and fashions from OSAA, IT groups achieve a clearer path to end-to-end architectures through which compute, information, brokers, and community safety align with widespread, open rules.
Black Hat Replace: The SAFE Tips and Increasing the Open Stack
Because the annual Black Hat USA convention kicks off, the Open Safe AI Alliance introduced a significant new ecosystem initiative.
The Linux Basis, in collaboration with alliance members, launched a Request for Feedback (RFC) on the Shared AI Findings Alternate (SAFE) pointers.
SAFE is designed to show agentic cybersecurity incidents and “close to misses” into shared safety throughout the enterprise. Reasonably than holding breach particulars locked behind non-disclosure agreements or confined to single-vendor ecosystems, SAFE supplies a framework to:
- Confidentially accumulate and analyze AI safety incidents, agent misbehaviors, and near-miss operational occasions.
- Inform impacted events rapidly whereas figuring out recurring management failures in actual time.
- Publish evidence-based working suggestions to decrease systemic danger for all defenders.
Alongside the SAFE framework, OSAA members are increasing the open software program layer. Nvidia, for example, launched Nvidia OpenShell, an open runtime that serves as an agent-level sandbox. OpenShell limits what an agent can see, contact, or execute, implementing express safety and privateness boundaries so autonomous instruments can’t entry techniques or information past their designated scope.
Collectively, these Black Hat bulletins reinforce a key precept for enterprise safety groups: securing agentic AI requires collective menace intelligence and open runtime guardrails, not simply remoted mannequin scanners.
Accelerating enterprise time to worth for AI
For enterprise IT, the query is much less “does this alliance matter?” and extra “how does it change my roadmap?” There are a number of direct impacts on time to worth:
- Quicker, safer experimentation: Open fashions and harnesses which might be inspectable and self-hostable let groups run high-risk safety workloads, similar to incident response or vulnerability discovery, on their very own infrastructure with out ready for closed suppliers to assist area of interest use circumstances.
- Diminished integration friction: An open protection stack spanning identification, format security, scanning, and safe improvement supplies a reference structure for AI safety, shortening design cycles and avoiding bespoke integrations for every device.
- Higher ROI from current fashions: Methods similar to typed I/O, pass-by-reference, and structured reminiscence can enhance accuracy and scale back token utilization with the identical fashions, turning stalled pilots into manufacturing deployments that meet each efficiency and finances constraints.
- Stronger systemic resilience: Open, multi-vendor ecosystems scale back the chance of single factors of failure or coverage shocks that come up when a small set of closed suppliers controls crucial defensive capability.
In follow, this implies AI safety tasks can shift from “hero-based” experiments to repeatable patterns. Groups can apply NOOA-style harness designs to their incident response brokers, code reviewers, or threat-hunting workflows, and combine these brokers with identification, logging, and scanning layers aligned with OSAA contributions.
Steerage for IT and safety leaders
For IT execs planning or scaling AI initiatives, a number of concrete steps might help them benefit from this:
- Deal with the alliance outputs as a reference structure. As OSAA parts mature, map them to your current stack: SPIFFE/SPIRE for agent identification, Safetensors for mannequin storage, Lightwell for supply-chain integrity, MDASH-like approaches for multi-model scanning, and harness patterns like NOOA on your agent platforms.
- Make harness engineering a crucial self-discipline. Don’t body agent tasks solely as “mannequin choice”; put money into harness design. That’s, use typed contracts, pass-by-reference, programmable loops, and express states, and measure their impression on accuracy, latency, and token prices.
- Align safety and improvement practices round brokers as code. With NOOA, brokers are Python courses that may be code-reviewed, unit-tested, and traced like some other crucial service. Lengthen CI/CD, safe coding, and observability practices to those harnesses relatively than treating them as one-off automation scripts.
- Advocate for open, inspectable defensive instruments in coverage and procurement. When evaluating AI safety merchandise or drafting inner insurance policies, emphasize assist for open fashions and harnesses that you would be able to run and examine on-prem or in your trusted cloud environments.
- Begin with high-leverage safety use circumstances. Early tasks may deal with vulnerability discovery in your individual codebase, automated configuration evaluation, incident triage, or threat-hunting brokers explicitly designed to work inside NOOA-style harnesses and OSAA-aligned identification and logging.
The Open Safe AI Alliance’s core thesis is that AI protection needn’t be opaque or centralized. With the suitable mixture of open fashions, rigorous harness engineering and shared defensive tooling, enterprises can construct AI techniques that aren’t solely highly effective but additionally observable, testable and below their management.
For IT leaders below stress to ship AI outcomes with out compromising safety, that may be a future price embracing now.
Different Information: Google is tightening Chrome extension insurance policies to curb abuse by limiting high-risk permissions, strengthening enterprise controls, and making it more durable for malicious add-ons to take advantage of browser entry.
